Graph-Encoded Imitation Learning for Optimization of Multi-Process Electric Load Shifting in the Steel Industry

Iron and steel production scenarios involve complex hierarchical relationships between processes, workshops, and production lines, as well as intricate process constraints. These factors lead to a large number of variables in the electric load shifting optimization (ELSO), making it challenging to solve efficiently using linear programming methods. Heuristic algorithms, while useful, often fail to ensure the feasibility of individuals within the population during the iterative process, which can result in solution instability. Moreover, traditional reinforcement learning methods, which require continuous interaction with the environment to optimize scheduling strategies, struggle to improve learning efficiency in this context. To address these challenges, this paper introduces expert-driven imitation learning (IL) into this problem. In addition, to capture the inherent characteristics of multi-process loads and the variations between them, we propose a multi-layer gated graph structure encoder, represented by nodes and undirected edges, to enhance the feature representation of the input loads. Finally, a multi-head attention (MHA) decoder is utilized to learn the scheduling strategies across multiple processes. Experimental results demonstrate that our proposed model converges more rapidly and exhibits greater solution stability for large-scale problems.